VULKAN CRASH FIX Gate DFlash graph argmax/top-k to CUDA backend

[scriptnapps]

FIX crashing on Vulkan windows build

DFlash builds ggml_argmax_ext / ggml_topk_ext into the draft graph, but only the CUDA backend implements their extended semantics; on other backends (Vulkan, Metal, CPU, ...) the stock argmax kernel leaves most of the compact ids/probs tensor uninitialized, producing garbage drafts. This PR adds a dflash_graph_argmax cparams flag, set only when the output device backend is CUDA, and falls back to full-logits CPU scanning (and full-logits verification, with a one-time warning) everywhere else. No behavior change on CUDA.

Additional information

Observed on AMD RX 9070 XT with the Vulkan backend, where DFlash drafting was broken out of the box. The check is name-based (strcmp(reg_name, "CUDA")); could be replaced with a proper backend capability query if other backends gain support.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: no